High precision in protein contact prediction using fully convolutional neural networks and minimal sequence features

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in genomics

Background:

• State-of-the-art protein contact prediction methods often integrate diverse sequence features beyond substitution frequencies.
• The necessity and contribution of these additional features for achieving high prediction accuracy remain unclear.
• Protein contact prediction is crucial for understanding protein structure and function.

Purpose of the Study:

• To investigate the sufficiency of simple protein sequence alignment statistics for state-of-the-art contact prediction.
• To develop and evaluate a deep neural network-based method (DeepCov) for protein contact prediction using only alignment data.
• To determine if complex, global statistical models are essential for precise contact predictions.

Main Methods:

• Utilized fully convolutional neural networks (CNNs) within the DeepCov framework.
• Employed amino-acid pair frequency or covariance data derived directly from protein sequence alignments as input.
• Avoided the use of global statistical methods like sparse inverse covariance or pseudolikelihood estimation.

Main Results:

• DeepCov achieved performance comparable to or exceeding existing methods like CCMpred and MetaPSICOV2.
• High prediction precision was maintained using relatively local sequence windows (around 15 residues).
• DeepCov demonstrated substantially improved precision over CCMpred and MetaPSICOV2 on shallow sequence alignments.

Conclusions:

• Simple alignment statistics, when processed by deep neural networks, are sufficient for state-of-the-art protein contact prediction.
• The DeepCov method offers a computationally efficient and accurate alternative, not requiring extensive feature engineering.
• These findings suggest that complex global models may not be strictly necessary for precise residue-residue contact predictions.